3/30/12 Performance of Iterative
Reconstruction Algorithms
AAPM Spring Clinical Meeting
March 17, 2012
Kirsten Boedeker, Rich Mather
Toshiba Medical Research Institute
Courtesy of Patrik Rogalla, UHN
CT overuse
Sodickson A, Radiology, 2009 251:175
1 3/30/12 Dose Reduction
!   Conventional Approaches
! mA Modulation, Gating
!   Post-processing algorithms
!   Bowtie
!   Active Collimation
Quantum Noise
!   Noise
!   Random background variations
!   Competes with true signal
!   Static on radio, “snow” on television
!   More photons
!   less noise
!   Signal to Noise ratio (SNR)
Quantum Noise
N = 225
25
75
150
Image noise is heavily dependent on the number of photons (quanta) that define it!
2 3/30/12 100,00
90,00
25,0
80,00
70,00
20,0
60,00
50,00
40,00
15,0
10,0
30,00
20,00
Exposure [mSv]
Image noise [%]
10%
20%
30%
40%
50%
60%
70%
80%
90%
0%noise
noise
5,0
10,00
0,00
0,0
25
50
75 100 125 150 175 200 225 250
Tube current [mAs]
Courtesy of Patrik Rogalla, UHN
Dose Reduction made easy!
100%?
75%
50%
25%
Courtesy of Patrik Rogalla
3 3/30/12 Image Reconstruction
!   Have raw data, need to make an image
!   2 major ways to do this
!   Fourier Backprojection
!   Iterative Reconstruction
!   Both older than dirt
!   Different strengths and tradeoffs for each
Acquiring Projection Data
red
e
t
l
Fi
Backprojection
4 3/30/12 Filtered Backprojection
5 3/30/12 FBP Advantages
!   Speed
!   50-60+ images per second recon
!   Well characterized
!   Primary recon since beginning of CT
!   Noise properties known; linear relationship between noise and
resolution
!   Familiar look and feel
!   Artifacts known
!   Linear
!   Predictable reconstruction behavior
FBP Disadvantages
!   Simplified assumptions
!   Point focal, point detector, pencil beam
!   Limits precision
!   Trouble with truncated data
!   Needs to know all the projections
!   Slight non-uniformities
!   Can be calibrated out
!   Direct tradeoff between noise and resolution can be limiting
Simple Iterative Example
13
11
7
10
17 18
1
8
9
8
18
15
11
15
Total variation = 12
8
13
6 3/30/12 Simple Iterative Example
13
14
6
11
17 18
Total variation = 7
3
5
8
9
16
11
11
15
8
13
Simple Iterative Example
13
13
8
10
18 18
Total variation = 0
3
5
8
11
15
13
11
15
8
13
Traditional Iterative Reconstruction
Forward
Project
(modeling)
Compare
~ equal
Update
Initial
guess
7 3/30/12 Modeling for IR
!   Statistical Modeling
!
!
!
!
Focused on controlling noise
Models only noise properties
Takes quantum noise into account
Does not improve resolution!
!   Physics modeling
!
!
!
!
Models all aspects of the scanner
Focal spot size, system geometry, beam energy, cone angle
Extremely complex- better the model, the better the image quality
Can improve both noise and resolution
Possible Iterative Advantages
!   Modeling
!   Allows more precise reconstructions
!   Can help with noise and resolution
!   Artifact reduction
!   Good with truncated datasets
!   Short scans
!   Cone beam
Traditional Iterative Disadvantages
!   Slow
!   Depending on model can be 400x slower
!   Complex
!   Modeling noise is relatively fast
!   Modeling physics is slow!
!   Non-linear
!   Can create plastic images
!   Poorly characterized
8 3/30/12 Vendor-specific Iterative Reconstruction
GE
Philips
Siemens
Toshiba
Algorithm
•  Veo
•  iDose4
•  SAFIRE
•  AIDR3D
•  Data domain
•  Data and Image
domain
•  Data and Image
domain
•  Data and Image
domain
GE’s iterative reconstruction- Veo
!   Traditional IR- Model based IR
!   Forward projection incorporates
!
!
!
!
!
!
Real focal spot
Real detector geometry
Cubic voxel
Broad beam
Statistical model of noise
Physics model
!   “Lower noise and higher resolution can be achieved within a
single image. ” – GE website
!   ~ 1 hour / case – Katsura et al, ECR March 2012
Philips’ iterative reconstruction- iDose4
!   Hybrid
!   Works in both raw data domain and image domain
!   Raw data
!   Targets noisy projections
!   Noisy data penalized, edges preserved
!   Image data
!   Noise model
!   Multi-frequency noise reduction
!   “The majority of factory protocols are reconstructed in 60
seconds or less ” – Philips website
9 3/30/12 Siemens’ iterative reconstruction- SAFIRE
!   Hybrid
!   Works in both raw data domain and image domain
!   Primary work done in image space
!   “Iteratively ‘cleaning up’ and removing image noise without degrading
image sharpness”
!   Periodic comparison to sinogram
!   Forward project into raw data domain
!   Compare with original acquisition data
!   “SAFIRE can achieve significant radiation dose reduction ” –
Siemens website
10 3/30/12 Adaptive Iterative Dose Reduction (AIDR3)
AIDR 3D* is the latest evolution and Toshiba’s 3rd generation of iterative reconstruction
technology that has been fully integrated in to the imaging chain. AIDR 3D* has been optimized
for a busy workflow and adds mere seconds to total reconstruction time.
Scanner
Model
AIDR Image
Anatomical Model
Based Optimization
Update
Object
Projection
Noise
Reduction
+
Statistical
Model
Blending %
Iterative
Adaptive
Dose Reduction – AIDR 3D
Standard dose FBP
Full Dose – 2mm sl
*Pending FDA Clearance
IR
AIDR (Std) – Low Dose Abdomen
Courtesy of Patrik Rogalla, UHN
11 3/30/12 Reduced dose FBP
IR
35 % Dose reduction
AIDR (Std) – Low Dose Abdomen
35% dose reduction
Courtesy of Patrik Rogalla, UHN
Reduced dose FBP
IR
50% dose reduction
Courtesy of Patrik Rogalla, UHN
Low dose FBP
IR
65% dose reduction
12 3/30/12 Low dose FBP
IR
75% dose reduction
Courtesy of Patrik Rogalla, UHN
Low dose FBP
IR
85% dose reduction
Courtesy of Patrik Rogalla, UHN
Image Quality
Medical:
Technical:
•  we can see very small structures
high contrast resolution
•  we can see subtle density differences
low contrast resolution
•  no motion artefacts
temporal resolution
•  low image pixel noise
pixel SD, noise power spectrum
•  sharp contours, crisp image
kernel, edge preservation
•  tissue contrast
tube settings, CM
Mental:
•  “nice images”
psychology
Courtesy of Patrik Rogalla, UHN
13 3/30/12 Evaluating Image Quality
!   Noise
!   Spatial Resolution
! Detectability
SD=21.5HU
SD=21.5HU
Both images have the same standard deviation
Impacts Detectability
14 3/30/12 Sagittal Plane
Axial Plane
Iterative Reconstruction
Texture Changes
15 3/30/12 Noise Power Spectrum
!   Based on Fourier technique, images of uniform, noise-only
material are converted into frequency space to yield a power
spectrum.
!   Shows in which spatial frequencies the noise power is
concentrated.
!   Area under NPS curve is equal to the variance.
Low Frequency NPS, Large Grain Noise
Smooth Filter
0.00025
HU2pix 2
0.0002
0.00015
Smooth
0.0001
0.00005
0
0
2
4
6
8
10
12
frequency (lp/cm)
High Frequency NPS, Fine Grain Noise
Sharp Filter
0.0045
0.004
0.0035
HU2pix 2
0.003
0.0025
Sharp
0.002
0.0015
0.001
0.0005
0
0
2
4
6
8
10
12
frequency (lp/cm)
16 3/30/12 Iterative Reconstruction and NPS
!   IR can shift the NPS to lower frequencies
!   Amount of the shift can depend on:
!   1. Dose level
!   2. Algorithm Strength
Marin D et al Radiology 2010
Dose = 3.4 mGy
150
FBP
IR1
IR2
NPS (HU2mm2)
100
IR3
50
0
0
0.2
0.4
0.6
0.8
Spatial frequency (mm-1)
Courtesy of Ehsan Samei
17 3/30/12 18
16
14
12
10
8
6
4
2
0
0
0.2
0.4
0.6
0.8
1
1.2
frequency (1/mm)
600
500
Amplitude
400
300
200
100
0
0
NPS (HU2mm2)
150
100
0.2
Dose = 3.4 mGy
FBP
IR1
IR2
IR3
50
0
0
0.4
150
100
0.6
0.8
frequency (1/mm)
Dose = 13.5 mGy
FBP
FBP
IR1
IR1
IR2
IR2
IR3
IR3
0
1.2
150
Dose = 27.0 mGy
FBP
IR1
IR2
IR3
100
50
50
0.2
0.4
0.6
0.8
Spatial frequency (mm-1)
1
0
0.2
0.4
0.6
0.8
Spatial frequency (mm-1)
0
0
0.2
0.4
0.6
0.8
Spatial frequency (mm-1)
Dose reduction and texture change of IR
18 3/30/12 Noise and Iterative Reconstruction
!   Like with FBP, standard deviation does not tell the whole story
!   NPS can vary with IR algorithm and strength
!   NPS can vary with dose level
!   SD and NPS should be quantified for range of typical use.
Resolution
MTF Bead or Wire
19 3/30/12 MTF
MTF
1
0.9
0.8
0.7
MTF
0.6
0.5
0.4
0.3
0.2
0.1
0
0
5
10
15
20
lp/cm
Slice Sensitivity Profile
Slice Sensitivity Profile
1.2
Amplitude
1
0.8
0.6
64-row Helical
0.4
160-row Ultrahelical
0.2
0
-0.2 0
0.5
1
1.5
2
2.5
Position (mm)
Modulation Transfer Function
!   Highly attenuating wire or bead for test object
!   Presumes linear behavior of algorithm
!   Linear algorithm =>Performance at high contrast reflects spatial
resolution properties at low contrasts
60
20 3/30/12 Iterative Reconstruction
!   Non Linear Algorithms spatial resolution preservation depends
on contrast level and noise level
!   Traditional test objects not robust
!   Traditional test condition (very small FOV or pre-sampled) do not reflect
clinical scanning/display conditions
Adaptive image filters
Edge Definition
1
Edge Extraction
0
k（0～1)
1-k
Image
Smoothing
Original
k
Enhancement
Non-linear Algorithms
Taken from : Okurura, et al,. New Method of Evaluating Edge-preserving Adaptive Filters for Computed Tomography (CT):
Digitial Phantom Method. Japanese Society of Radiological Technology. 2006
21 3/30/12 TTF measurements:
Task-specific, edge technique
•  Edge of rods. Similar to
MTF measurements, but
Air
Teflon
Acrylic
•  Task-specific:
object contrast, dose,
and recon
Polystyrene
Courtesy
Ehsan Samei
Low contrast, high dose
High contrast, high dose
1
1
FBP
FBP
IR1
IR2
0.8
IR1
IR2
0.8
IR3
IR3
0.6
TTF
TTF
0.6
0.4
0.4
0.2
0.2
0
0
0.2
0.4
0.6
Spatial frequency (mm-1)
0
0.8
0
0.2
0.4
0.6
Spatial frequency (mm-1)
1
1
FBP
FBP
IR1
IR2
0.8
IR1
IR2
0.8
IR3
IR3
0.6
TTF
TTF
0.6
0.4
Courtesy
Ehsan
Samei
0.8
Low contrast, low dose
High contrast, low dose
0.4
0.2
0
0.2
0
0.2
0.4
0.6
Spatial frequency (mm-1)
0.8
0
0
0.2
0.4
0.6
Spatial frequency (mm-1)
0.8
Detectability and Image Quality
!   What is low dose?
!   Full Dose w FBP vs Reduced Dose w IR
22 3/30/12 Courtesy of Patrik Rogalla, UHN
Dose
Image Quality
Diagnostic information comes from dose!
Diagnostic information comes from dose
23 3/30/12 Simulated dose: 0.4 mAs
van Gelder et al. Radiology 2004
Courtesy of Patrik Rogalla, UHN
50 mAs
25 mAs
6.3 mAs
1.6 mAs
From:
Dr. Jaap Stooker
The Netherlands
Courtesy of Patrik Rogalla, UHN
poor low contrast detectability
FBP 0.5 mm slice
Courtesy
IR 0.5 mm slice
Patrik Rogalla UHN
24 3/30/12 Courtesy of
Ehsan
Samei
73
3/30/12
!   How do we quantify dose reduction associated with IR?
!   Objective phantom data
!   Reproducible
!   The focus is on the detection tasks -> most challenging in low
dose imaging conditions is Low Contrast Detectability.
Observer study
B. Imaging
Task / Phantom
Choice
(Lesion contrast /
size, background)
A. Observer
(Human / Model)
C. Protocols &
dose levels to
assess
(Clinical
relevance)
D. Study
(AFC, ROC…)
Courtesy MITA CT Physics
Committee
75
25 3/30/12 Choose an observer
Human observer
Model observer
Positives
Challenges
Human Observers
Model Observers
•  Straightforward
implementation.
•  Objective and consistent
•  Directly incorporates human
perception
•  Demonstrated correlation with
human performance for certain
tasks
•  Time consuming – observer
fatigue – statistical power
•  How choose from the wide variety
of published observers?
•  Controlled study needed
•  Need validation with human
observers for CT non-linear algos
•  Inter- and Intra-Observer
variability
77
•  No single model observers can do
all the tasks (detection,
estimation, etc)
Courtesy MITA CT Physics
Committee
Model Observers
Choice of model observer
•  Many observers to choose from
•
•
•
•
•
Ideal Observer
Hotelling
Non-prewhitening (NPW)
Channelized models
Eye and Internal Noise filters
NPW =
∫ O( f )
2
MTF ( f ) 2 df
2
2
2
∫ O( f ) MTF ( f ) NPS ( f ) df
Each observer represents a different set of starting assumptions.
The ideal observer, for example, assumes the correlations in
the noise can be undone. The non-prewhitening model
assumes they cannot (bc the human eye cannot undo them).
The NPWE (with Eye filter) incorporates the frequency response
of the human eye.
26 3/30/12 Courtesy of Ehsan Samei
Observer Study Design: Imaging Task
!   Type of Task
81
!   Classification task?
!   Estimation task?
27 3/30/12 Observer Study Design: Imaging Task
82
1.  Defining the test object (i.e. signal)
•
What is the object of interest?
•
•
•
•
•
What is background of interest?
•
•
•
!
Sphere? Simulated anatomy?
Contrast level
Size
Position in field
Correlated electronic and quantum noise (water phantom)
Anatomical noise
SKE/BKE? Search?
Known Object (SKE) & Location, BKE
!
Known Object (SKE), Unknown Location, BKE
Possible Object
Locations
!
Shape discrimination and size
estimation
Courtesy of MITA CT Physics Group
Positives
Industry Standard
Phantoms (Catphan,
ACR phantom, etc.)
Custom Phantoms
•  Standardized
•  Can be tailored to task
•  Reproducible
Challenges
•  Fixed object sizes and
contrasts
•  Limited ability to isolate
and analyze individual
objects
•  Non-standard
•  Not readily available to the
field
•  Need to be defined
84
Courtesy of MITA CT Physics Group
28 3/30/12 Imaging Task
!   Phantom/Task
Ø  Low contrast detection task for different contrast levels: 0.3%, 0.5%, 1.0%
Ø  Selectable disk sizes (9 cylinders of diameters): 2, 3, 4, 5, 6, 7, 8, 9, 15 mm
!   Positives
Ø  Industry standard phantom
Ø  Radial symmetry in object locations
Ø  Background closer to soft tissue
Ø  Relatively higher flexibility in reference dose selection due to availability of multiple contrast and disk
sizes
!   Challenges
Ø
Ø
Ø
Ø
Ø
Limited ability to ioslate and analyze individual test objects
Single realization for each object (contrast & size)
Available disk sizes and contrast may not be enough to cover range of all dose levels/protocols
Known pattern may reduce imaging task to “Signal Known Exactly”
Inter-phantom variability due to tolerance (some Catphans better than others)
85
Courtesy of MITA CT Physics Group
Imaging Task
!   Phantom/Task
Ø  Low contrast detection task for 0.6% contrast (fixed)
Ø  Selectable disk sizes out of five cylindrical objects (discrete)
!   Positives
Ø  Industry standard phantom
Ø  Better spacing processing
Ø  Multiple realizations of the same object (4)
!   Challenges
Ø  Requires flexibility in reference dose selection/protocol due to fixed contrast and discrete
diameters.
Ø  Compromise from radial symmetry in object locations.
Ø  Background denser than soft tissue -- clinical relevance ?
Ø  Not commonly used outside the US.
Ø  Known pattern may reduce imaging task to “Signal Known Exactly”
86
Courtesy of MITA CT Physics Group
Imaging Task
!   For non-linear processes (where performance varies nonlinearly with contrast, position, dose, etc) what protocols of
interest capture a good representation of performance?
!   Typical Performance (i.e. Clinical protocols)
!   Max Performance
!   Produce non-trivial comparisons in a dose range typical for
the organ (e.g. a non-trivial ROC curve)
!   The choice of imaging task should result in clinically relevant dose
levels.
!   Reproducible with commercially available phantoms in the
field?
29 3/30/12 Study Design
!   ROC Study
88
!   Traditional method with a “confidence” rating scale.
!   More time consuming to run, requires greater observer expertise.
!   Many options: LROC, FROC, etc
Study Design
!   Alternative Forced Choice (AFC)
!   2-AFC experiments are the easiest / fastest to run.
!   4-AFC experiments are not much more difficult, and provide better statistics for
reasonable sample sizes (~100 images).
Figure of Merit
-AUC
-d’
-SNR
-Percent Correct
30 3/30/12 1.  Model Observer
2.  0.5% rods Catphan
at 50mA vs 150mA
3.  2AFC
How does PC and AUC
compare?
Template
10
20
30
40
50
60
0
10
20
30
40
50
60
70
Sample Signal Present Image
10
20
10
20
Value 1
30
30
40
40
50
50
60
0
10
20
30
40
50
60
70
60
0
10
20
30
40
50
60
70
31 3/30/12 Sample Absent Present Image
10
20
10
20
Value 2
30
30
40
40
50
50
60
0
10
20
30
40
50
60
70
60
0
10
20
30
40
50
60
70
Who won?
!   If Value1 (signal present) exceeds Value2 (signal absent), then
a correct “hit” is recorded.
!   If Value2 exceeds Value1 a “miss” is recorded.
!   Process is repeated for a large number of signal present and
signal absent images
32 3/30/12 AUC vs. dose
1
AUC
0.9
0.8
Dose
reduction
0.7
FBP
IR1
IR2
IR3
0.6
Courtesy of
Ehsan Samei
0.5
0
5
10
15
20
CTDIvol(mGy)
25
30
•  Medium patient
•  10 mm lesion diameter, 250HU lesion enhancement
Courtesy of Ehsan Samei
Conclusions
!   Goal = Dose Reduction
!   Iterative Reconstruction offers excellent potential dose
reduction and good noise/resolution properties
!
!
!
!
Slow
“Unnatural” look and feel
Some loss of edge/detail
NON-LINEAR
!   IQ Characterization
!   Traditional Metrics come up short
!   NPS at variety of dose levels and IR strengths
!   Contrast-dependent MTF
! Detectability Studies
33